Вы находитесь на странице: 1из 12

Electoral Studies 21 (2002) 3546


The endogeneity problem in electoral studies: a

critical re-examination of Duvergers
mechanical effect
Kenneth Benoit

Department of Political Science, Trinity College, University of Dublin, Dublin, Ireland

Studies of electoral law consequences typically treat electoral laws as exogenous factors
affecting political party systems, even while acknowledging that political parties often tailor
electoral institutions to suit their own distributional needs. This study represents a departure
from that approach, directly examining one aspect of the endogeneity of electoral systems:
the endogeneity of Duvergers mechanical effect. Theory clearly posits that the Duvergerian
psychological effect of electoral rules occurs in anticipation of their reductive mechanical
effect, yet in empirical models this endogenous character is typically ignored. In this paper I
formalize the two types of Duvergerian effects of electoral laws in a structural model and
implement this model using two-stage-least-squares regression to re-estimate the mechanical
effect model of Amorim Neto and Cox [Electoral institutions, cleavage structures, and the
number of parties. Am. J. Polit. Sci. 41 (1997) 149-174] and Cox [Making Votes Count:
Strategic Communication in the Worlds Electoral Systems. Cambridge University Press
(1997)]. I also generalize the model and compare it to two other approaches taken by Ordeshook and Shvetsova [Ethnic heterogeneity, district magnitude, and the number of parties. Am.
J. Polit. Sci. 38 (1994) 100-123] and Taagepera and Shugart [Predicting the number of parties:
a quantitative model of Duvergers mechanical effect. Am. Polit. Sci. Rev. 87 (1993) 455464]. The results indicate that because electoral structure affects the number of parties in
the legislature both directly through the mechanical effect as well as indirectly through the
psychological effect, simple OLS estimates that do not take into account this endogenous

Previously presented at the 1999 Annual Meeting of the American Political Science Association,
Atlanta Mariott Marquis and Atlanta Hilton and Towers, September 25, 1999. A full replication dataset
is available upon request from the author. The author thanks Gary Cox and Mike Harrison for ideas and
comments and Matthew Shugart for assistance with his original dataset.
* Tel.: +353-1-608-2941; fax: +353-1-677-0546.
E-mail address: kbenoit@tcd.ie (K. Benoit).

0261-3794/01/$ - see front matter 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 2 6 1 - 3 7 9 4 ( 0 0 ) 0 0 0 3 3 - 0


K. Benoit / Electoral Studies 21 (2002) 3546

model will overestimate the mechanical effect by 45100%. 2001 Elsevier Science Ltd. All
rights reserved.
Keywords: Electoral systems; Mechanical effect; Psychological effect; Endogeneity; Duverger

1. Two effects of electoral laws

Duvergers famous proposition that the simple-majority single-ballot system
favors the two-party system (Duverger, 1951, p. 217) has generated a huge field
of subsequent empirical research linking electoral structure to the number of political
parties present in a nations political system (Amorim Neto and Cox, 1997; Ordeshook and Shvetsova, 1994; Lijphart, 1994; Taagepera and Shugart 1989, 1993; Blais
and Carty, 1991; Rae, 1967). Theory identifies a dual mechanism by which electoral
institutions shape political party systems: two forces working together: a mechanical
and a psychological factor (Duverger, 1951, p. 224). The mechanical effect of electoral systems describes how the electoral rules constrain the seats that can be awarded
from distributions of votes, while the psychological effect deals with the shaping
of party and voter strategies in anticipation of the electoral functions mechanical
constraints. Analysis of the mechanical effect considers the number of parties winning seats as a dependent variable, using electoral structuretypically represented
by district magnitudeas the key explanatory variable. Research into the psychological effect, on the other hand, focuses on the role of electoral rules in shaping the
number of parties contesting seats, as well as the way that votes for these parties
are cast, often controlling for such factors as ethnic cleavages (Ordeshook and Shvetsova, 1994), issue dimensions (Taagepera and Grofman, 1985), and the character
and timing of presidential elections (Amorim Neto and Cox, 1997).
Empirical research since Duverger, however, has not always drawn a rigid distinction between the two types of effects (for an explanation see Blais and Carty, 1991),
nor has it fully explored the theoretical implications of these two mechanisms for
empirical modeling. Why should it be important to distinguish between the two
effects and to estimate each with accuracy? First, without a clear distinction, as this
paper demonstrates, estimation of the mechanical effect in particular tends to be
biased. Since the psychological effect consists of anticipations of the mechanical
effect, then an incorrect understanding of the mechanical effect may also warp the
psychological effect. Second, real-world political actors, especially those considering
alternative electoral systems, require an accurate understanding of the mechanical
consequences of various electoral rules. Especially in cases of initial electoral system
choice where configurations of political parties may be shaped prior to institutions,
decision-makers will be interested in how electoral rules will reduce the number of
parties through purely mechanical means. Since political science knowledge increasingly influences decision-makers selecting among alternative electoral institutions, a
clear and accurate understanding of the mechanical properties of these institutions
is of critical importance.
In this study I focus renewed attention on the estimation of the mechanical effect.

K. Benoit / Electoral Studies 21 (2002) 3546


In what follows I express the mechanical and psychological effects in a single structural model, then apply this model to replicate and re-estimate the mechanical effect
reported in Amorim Neto and Cox (1997) and Cox (1997).1 The results indicate that
because electoral structure psychologically influences the composition of votes which
condition the mechanical effect, non-structural estimates of electoral structures
mechanical effect will be overestimated. Variations of this result applied to two other
approaches to estimating the mechanical effect taken from Ordeshook and Shvetsova
(1994) and Taagepera and Shugart (1993) reveal a similar upward bias.

2. Data
My purpose here is to offer a clarification and correction of previous models of
the mechanical effect, not to extend them using essentially new data. For this reason
I have chosen to adhere closely to the data and design of previous empirical research
into the mechanical effect. While this approach carries over all of the flaws in the
data, variable names, measures, and estimation methods of the original studies, it
has the advantage of isolating the essential model for direct comparison. Substantive
work to follow will hopefully incorporate the lessons derived from this re-examination, although a full implementation of these lessons is left to future work. For
example, mechanical effects as defined here really operate at the district level, and
therefore are likely to be obscured when observing only aggregate results.2 Since
the objective of this re-analysis is to focus on an issue in modeling the mechanical
effect, however, the data and methods used here closely follow those of Cox (1997)
and Amorim Neto and Cox (1997) for purposes of directly comparing and extending
previous estimates of the mechanical effect.
The dataset from Amorim Neto and Cox (1997) provides 54 cross-national observations of elections from the 1980s. The dependent variable ENPS refers to the
effective number of legislative parties (1/s2i , where si is party is seat share). The
independent variables include ENPV, the effective number of parties in the electorate
(analogous to ENPS but using vi as party is vote share); UPPER, indicating the
proportion of all assembly seats allocated in the upper tier(s) of the electoral system;
and LML, the natural logarithm of the median legislators district magnitude. Using
the logarithm of district magnitude M is standard practice, designed to indicate the
marginally decreasing consequences of M on the number of parties; for the purposes
of replicating Amorim Neto and Cox (1997) I follow their use of the natural logarithm. Finally, ENETH represents the effective number of electoral groups (analogous

Controlling for the number of parties in the electorate was also the approach of Blais and Carty
(1991), to which the following discussion also applies.
Another issue which might be raised concerns heteroskedasticity. While heteroskedasticity may be
present in the data, I did not report heteroskedasticity corrections of the standard errors for purposes of
adhering to the previous models. Testing revealed, however that adding these corrections does not change
the significance of any of the estimated coefficients.


K. Benoit / Electoral Studies 21 (2002) 3546

to the party counts but using gi to denote the proportion of the population of ethnic
group i), based on data collected by Ordeshook and Shvetsova (1994).

3. The model
The mechanical effect of electoral rules may be thought of as a physical limit on
the number of parties which can gain representation. For instance, in any electoral
district at most M parties may win seats (where M represents district magnitude),
no matter how many parties contest the M seats.3 The number of parties winning
seats when M1, however, will depend on the number of parties receiving votes as
well as their relative shares of the votes. Theoretically this is the complete information set necessary to observe the mechanical effect: the electoral rules and the
votes which the rules convert into seats. Once electoral rules are fixed, their effect
is mechanical in the sense that no human manipulation or strategy is involved
(Taagepera and Shugart, 1989). It is mechanical in the sense of accounting not for
political issues, personalities, or culture but only for the limits imposed on such
features by institutional structures (Taagepera and Shugart, 1993, p. 456). Eq. (1)
states this relationship formally, specifying b1 and b2 as the structural parameters
indicating the mechanical effect (and using the variable notation from Amorim Neto
and Cox, 1997).




Eq. (2) specifies the psychological effect, modeling the effective number of parties
receiving votes as a function of electoral rules as well as exogenous social factors
such as ethnic cleavages. It also highlights the endogenous nature of ENPV in Eq.
(2), since the psychological effect is driven by the perceived workings of the mechanical effect in Eq. (1). ENPV is also mechanically related to ENPS, since ENPS will
be equal to or less than ENPV (unless an electoral rule awards seats to parties receiving zero votes or the difference is an artifact of aggregation from districts).
Together the equations fully specify the structural relationship between the mechanical and psychological effects of electoral rules and the additional influence of
sociological factors on influencing the number of parties that both contest and win
office. The first equation separates the filtering of votes into seats by electoral rules,
for a given set of votes. The second equation describes how this set of votes is

This is what Taagepera and Shugart (1993, p. 458) forebodingly call the Forbidden Zone, placing
an upper limit of 1.0 on a linear coefficient relating ML to ENPS. While this upper limit of 1.0 will not
apply to the coefficient of LML since LML is logged, in real data the estimate will tends almost always
less than 1.0leading to estimates in practice of 01.0 even for the logged measures of district magnitude.
It is also possible to construct exceptions when district-level data is aggregated, although none exist in
the data used in this study.

K. Benoit / Electoral Studies 21 (2002) 3546


determined, including its shaping by the the anticipations, both by elites and voters,
of the workings of the mechanical factor (Blais and Carty, 1991, p. 92). Electoral
rules therefore exert a double effect: first in influencing the number of parties that
compete and the concentration of votes they receive, and second by controlling the
conversion of these parties votes into seats. In non-mathematical terms, votes which
are converted by the mechanical effect into seats have been pre-filtered by the
psychological effect, by voters and parties having some knowledge of how the rules
affect the chances of winning seats. It also implies that any observation we make
of the mechanical effect of electoral rules will also represent the influence of the
psychological effect, unless we have a method of taking into account the structural
relationship. While this endogenous relationship is sometimes implicit to previous
models of Duvergers effects, none have attempted to control for it when making
empirical estimates. The model here is therefore distinguished from previous characterizations by the explicitness of its structural treatment of the mechanical and
psychological effects and its direct control for the consequences of this endogenous relationship.
Pictorially we can represent Eqs. (1) and (2) in a path model (Fig. 1), highlighting
the means by which electoral institutions affect seats both directly (the mechanical
effect) and indirectly (the psychological effect) by shaping the votes to be converted
into seats. Because these functions are modeled as stochastic, the error terms U1 and
U2 also affect both models.4 In terms of Eqs. (1) and (2), the structural parameters

Fig. 1.

A path model of Duvergers two effects.

If we strictly interpret the views of Taagepera and Shugart (1993), then the mechanical effect is not
really a stochastic function at all. In practice, however, it is modeled with a residual term, since summary


K. Benoit / Electoral Studies 21 (2002) 3546

b1 and b2 represent the mechanical effect, while g1 and g2 represent the psychological effect.
The proper method of estimating these parameters depends on the structural
assumptions we can make about the correlation of error terms U1 and U2. If these
error terms are correlated, then in the language of structural equations, the model is
non-recursive. The consequences are that OLS estimates of the structural parameters
will be both biased and inconsistent. In the case of the two effects of electoral systems represented by Eqs. (1) and (2), there is excellent reason to believe that the
errors terms are correlated. First, because the tendency to reduce the number of
parties in both the mechanical and the psychological effects is driven by the same
set of institutional causes, any observation-specific forcessuch those uniquely
occurring in one country or electionnot captured in the systematic part of the
model are likely to have a similar affect on both error terms. For example, an idiosyncratic election where the number of parties competing for election was unusually
fragmented would also result in a potentially fragmented legislature, with the deviation from the psychological effect reflected directly as a deviation in the mechanical
effect. In addition, measurement problems are likely to affect both the psychological
and the mechanical models in a similar way. This is because electoral statistics such
as formula and district magnitude are invariably classified into discrete categories
or averaged at the national level, introducing potentially correlated disturbances
between the psychological and mechanical effect models which both incorporate
them. Both sets of forces should lead us to suspect a structural correlation and to
attempt modeling the problem using a more appropriate estimator. Two-stages leastsquares (2SLS) is one of the simplest methods of estimating structural parameters
for hierarchical and non-recursive models of this sort, producing consistent estimates
of structural parameters (Greene, 1993, pp. 6024). In the two sections which follow
I replicate and then re-estimate using 2SLS three previous empirical models of the
mechanical effect.

4. Re-estimating Amorim Neto and Cox (1997)

Amorim Neto and Coxs (1997) estimates of the mechanical effect provide a suitable setting for testing the model and its implications for empirical observations of
the mechanical effect. First, Amorim Neto and Coxs (1997) study, also reproduced
in Cox (1997), is familiar to many electoral systems researchers and uses variables
and methods common to most previous empirical studies of electoral system consequences. Second, the precise nature of the original study and the thorough description
of its data and variables facilitates replication. Finally, the study also estimated the
characterizations of electoral structure (e.g. district magnitude) are almost never complete, since data is
national- and not district-level, and since each observation of ENPV itself contains error. The empirical
estimation of the mechanical effect can therefore be thought of as an average effect for the system.
While theorizing the mechanical effect to be deterministic raises interesting implications, its treatment
for estimation purposes as stochastic here is completely consistent with prior practice.

K. Benoit / Electoral Studies 21 (2002) 3546


psychological effect, allowing for an extension of the original results using the structural model developed here.
Result (1) in Table 1 replicates Amorim Neto and Coxs (1997, Table 1) estimate
of the mechanical effect, using ENPV, ENPV*LML, and ENPV*UPPER as independent variables. Their study also included estimates of the psychological effect,
providing a means to obtain predicted values for ENPV to use in a 2SLS procedure
to re-estimate Result (1).5 This yielded consistent estimates for the effect of the
electoral institutions in Result (2) of Table 1indicating that the effect of the multiplicative LML variable is over-estimated by OLS by more than 45% relative to the
2SLS result (0.080 versus 0.055). Note that this replication did not involve any new
variables or any respecification of the original Amorim Neto and Cox (1997)
Table 1
Modeling the mechanical effect. Dependent variable: ENPSa


Adj. R2

Amorim Neto
and Cox (1997,
Table 1)


No interactive

Ordeshook and
(1994) approach

Taagepera and
Shugart (1993)















All data are from Amorim Neto and Cox (1997). Standard errors are in parentheses.

The estimate of the psychological effect I used was [EN
5.95PROXIMITY+2.14PROXIMITYENPRES+.51ENETH, all statistically significant at the p0.05
level, which I replicated in order to obtain predicted values for ENPV. The variables PROXIMITY and
ENPRES refer to the time between legislative and presidential elections and to the effective number of
presidential candidates. Although the inclusion by Amorim Neto and Cox (1997) of presidential variables
is unique, for replication purposes I retain their specification here. One reason not to include variables
on the psychological effect on presidential candidate is that these may also be determined by some of
the exogenous social factors included as control variables.


K. Benoit / Electoral Studies 21 (2002) 3546

regressions, but instead simply combined information from the separate estimates of
each effect into a structural model.
How should these results be interpreted? The answer has to do with the difference
between the total effect of district magnitude on the number of elected parties versus
the specific mechanical effect caused by electoral structure. The total effect of district
magnitude, for instance, includes both b1 and g1, but the mechanical effect refers
only to the former. When the votes to be filtered through the electoral rules have
already been pre-filtered in anticipation of the actual mechanical process, then the
non-structural model provides estimates that are too large, identifying a mechanical
effect from what is actually a total effect.
This endogeneity of the psychological effect can be seen when we compare ENPV
and ENPS for systems even with very different median district magnitudes (Table
2). For some of the elections from the Amorim Neto and Cox (1997) dataset using
extremely disproportional first-past-the-post systems we would observe, because of
the prefiltering of the psychological effect, results just as proportional as those from
systems with PR rules and much higher district magnitudes. The United States, for
example, using a single-member-district system with plurality rules, has 2.03 effective parties in the electorate and 1.95 parties winning seats, for a ratio of 1.04. Similar
results for single-member-district systems are the Bahamas and St Kitts and Nevis
with ratios of 1.08 and 1.00 respectively. Nearly identical ratios are also observed
in systems with much larger district magnitudes, such as Malta (M=5), and in Portugal, Finland, and Bolivia, with magnitudes between 16 and 17.5. In each case, the
psychological effect causes an equilibration between the number of parties that contest elections and the number that win them, and this effect is quite independent of
the mechanical effect of the electoral rules. These examples are selected of course
the full range can be assessed from the regression resultsbut they serve as illustrations of the dangers of observing mechanical effects in situations where the
psychological effect has already caused the parties competing and the parties winning
to equilibrate. This should deter us from the conclusion that the proper formulation
is one which ENPS would equal ENPV, were the electoral system perfectly proportional, with strong electoral systems reducing ENPS below ENPV (Amorim Neto
Table 2
Examples of psychological effect of pre-filteringa
United States
St Kitts and Nevis







Median M

All data are from Amorim Neto and Cox (1997). None of these examples have additional elements
such as upper-level compensation lists or additional-member systems.

K. Benoit / Electoral Studies 21 (2002) 3546


and Cox, 1997, pp. 1612) to the extent that the psychological effect does not fully
equilibrate ENPV.
Result (3) from Table 1 estimates a model of the mechanical effect without
multiplying all of the other variables by ENPV, but retaining ENPV as a standard
control variable. This simple additive formulation is consistent with previous
models of the mechanical effect and facilitates direct comparison for the purposes
of illustrating the bias caused by estimating the mechanical effect separately. The
coefficient of UPPER is not significant when not multiplied by ENPV, but for
purposes of comparison I keep it and simply focus the remaining discussion on
district magnitude (LML). The result of 0.253 as a coefficient on LML may be
interpreted as the true mechanical effect estimate of structural parameter b1 from
Eq. (1). This does not mean that the total consequences of district magnitude on
the effective number of parties is 0.253, but merely that this is a better measure
of the strictly mechanical effect of district magnitude where ENPV is fixed and
exogenous. The next section compares this figure to two other previous
approaches to estimating the mechanical effect.

5. Comparison to other models

Comparison of the structural approach to modeling the mechanical effect may be
extended to two other classes of previous models. The first type, included in Ordeshook and Shvetsova (1994), includes sociological control variables directly in the
model of the mechanical effect. The second, outlined (although not estimated
empirically) by Taagepera and Shugart (1993), disregards control variables altogether. This section demonstrates the superiority of the structural model to both of
these approaches.6
Result (4) in Table 1 uses the specification of Ordeshook and Shvetsova (1994)
with the data from Amorim Neto and Cox (1997).7 In this specification the operation of the mechanical effect is assessed while holding sociological factors constant. Yet because sociological factors do not intrude on the mechanical process
of vote-to-seat conversion, ENETH is really just a proxy for a measure of votes.
It is incomplete, however, and still results in an upwards-biased estimate of b1
as compared to Result (3)in fact it more than doubles the estimate of the mechanical effect. This occurs because ENETH is only a partial determinant of ENPV,

This section does not strictly replicate either of these two studies, although the Amorim Neto and
Cox (1997) dataset is essentially the same as that used by Ordeshook and Shvetsova (1994). I was not
able to use Taagepera and Shugarts (1993) dataset because even with Matthew Shugarts willing help I
was unable to reconstruct it exactly, and because their estimation procedure is not an empirical one which
can be replicated in the same way. Hence I replicate their model specifications while using the Amorim
Neto and Cox (1997) dataset, which keeps attention centered on the modeling issues which are the focus
of this paper.
In fact the data are essentially the same, since Amorim Neto and Cox (1997) used the measure of
ENETH as provided by Shvetsova.


K. Benoit / Electoral Studies 21 (2002) 3546

and because the psychological effect of LML is still being conflated into the
mechanical estimate through the omission of ENPV. As a consequence the coefficient on LML in Result (4) is an estimate of a reduced form parameter combining several effects, rather than of the structural parameter b1. Appendix A demonstrates this result mathematically.
The second alternative approach to estimating the mechanical effect is simply
to estimate the effect of electoral institutions on the number of parties winning
seats without any control variables. This approach was taken by Taagepera and
Shugart (1993), and followed using the Amorim Neto and Cox (1997) dataset in
the last column of Table 1. It causes a similar upward bias in the estimate of b1,
ascribing a much greater mechanical consequence to district magnitude only for
the limits imposed... by institutional structures (Taagepera and Shugart, 1993,
p. 456), because it fails to control for any other influences. The results in Table
1 show likewise that the estimate of the electoral institution-only model yields
a coefficient double that estimated by the structural model. Here the problem of
endogeneity can be demonstrated as one of simple omitted variable bias: the
omitted variable ENPV is highly correlated with the included variable district
magnitude through the very nature of the psychological effect. In this case the
omitted variable ENPV is both positively correlated with ENPS and with the
included variable LML, causing a positive upward bias in the estimate of LMLs
mechanical effect, detailed mathematically in Appendix B. Because the approach
using ENETH also suffered from this omitted variable bias, it is not surprising
that the two estimates of 0.518 and 0.510 should be similar.

6. Concluding remarks
This study has focused on empirical models of the mechanical effect, attempting to
bring empirical estimates closer to the purely mechanical factor identified in electoral
systems theory. The results demonstrate quite clearly in the context of previous
research that non-structural empirical estimates of Duvergers mechanical effect will
be biased, yielding results stronger than those caused by the strictly mechanical operation of electoral rules. This bias occurs because the operation of electoral systems
on the number of parties winning seats operates twice, both on the conversion of
votes to seats but also on the composition of votes itself. This result is robust to
several alternative approaches to estimating the mechanical effect. A more correct
approach is to model the structural relationship using two-stage least-squares (e.g.)
to separate the purely mechanical workings of electoral rules.
An unbiased estimate of the strictly mechanical effect might be of particular
interest in situations where the number and composition of political parties might
be at least partially independent of the electoral rules. Such a situation is common
in emerging democracies who are selecting from alternative rules based on their
understanding of the mechanical effects of the alternatives. The separate estimation points to the fact that in equilibrium, where ENPV is prefiltered by
electoral constraints, the results might be proportional even though the electoral

K. Benoit / Electoral Studies 21 (2002) 3546


institutions were having an indisputable effect on the number of political parties

in the parliament. In such contexts it is the number of political parties in parliament that is the most important immediate issue to electoral law designers, rather
than the more theoretical issues pertaining to strategic voting and party exit and
entry which political scientists prefer to examine under the rubric of the psychological effect. The structural model estimates presented here provide for such
contexts a more accurate picture of the institutional effects as separated from the
equilibrium effects which may only occur after a learning process has taken place.
For political scientists interested in the Duvergerian effects of electoral law, the
approach outlined here clarifies our understanding of the mechanical effect and
focuses attention where it is probably more deserved: on the psychological effect
of electoral laws on political parties and voters.

Appendix A. Demonstration of bias using ENETH as a control variable

Substituting Eq. (2) into Eq. (1), we get:



where the last step substitutes reduced form parameters p0 through p3 for the structural parameters. This formulation highlights the relationship
since the effect of both ENPV and LML on ENPS are always positive. Ordeshook
and Shvetsovas (1994) procedure is equivalent to estimating p1, while the 2SLS
model provides a consistent estimate of b1. The estimate of p1 will always be greater
by the product of the effect b3 of ENPV in the mechanical effect model and the
effect g1 of LML in the psychological effect model.

Appendix B. Demonstration of bias using no control variables

Consider two reduced versions of Eq. (1), omitting the constant and UPPER for
simplicity (not affecting the results):
The OLS estimator d1 for d1 will be:


K. Benoit / Electoral Studies 21 (2002) 3546






LML2i b3







where the last term in Eq. (4) is equivalent to the slope coefficient from an OLS
regression of ENPV on LMLa measure of the psychological effect, in other words.
The estimates of the effect of LML in Eq. (4) will therefore be larger than the
estimate of b1 in Eq. (4) by the product of the magnitude of LML in the psychological
effect and by the direct correlation of ENPV with ENPS.

Amorim Neto, O., Cox, G.W., 1997. Electoral institutions, cleavage structures, and the number of parties.
American Journal of Political Science 41 (1), 149174.
Blais, A., Carty, R.K., 1991. The psychological impact of electoral laws: measuring Duvergers elusive
factor. British Journal of Political Science 21, 7993.
Cox, G., 1997. Making Votes Count: Strategic Coordination in the Worlds Electoral Systems. Cambridge
University Press, Cambridge, UK.
Duverger, M., 1951. Political Parties: Their Organization and Activity in the Modern State. Wiley,
New York.
Greene, W.H., 1993. Econometric Analysis, 2nd ed. Macmillan, New York.
Lijphart, A., 1994. A Study of Twenty-Seven Democracies 19451990. Oxford University Press, New
Ordeshook, P.C., Shvetsova, O.V., 1994. Ethnic heterogeneity, district magnitude, and the number of
parties. American Journal of Political Science 38 (1), 100123.
Rae, D.W., 1967. The Political Consequences of Electoral Laws. Yale University Press, New Haven.
Taagepera, R., Grofman, B., 1985. Rethinking Duvergers Law: predicting the effective number of parties
in plurality and PR systemsparties minus issues equals one. European Journal of Political Research
13, 341353.
Taagepera, R., Shugart, S., 1989. Seats and Votes: The Effects and Determinants of Electoral Systems.
Yale University Press, New Haven.
Taagepera, R., Shugart, M.S., 1993. Predicting the number of parties: a quantitative model of Duvergers
mechanical effect. American Political Science Review 87 (2), 455464.